A story you may have already heard
In April of this year, a software company called PocketOS lost its entire production database in nine seconds.
An AI coding agent had been asked to fix a routine credential mismatch. Instead of flagging the problem for a human to review, the agent made its own decision. It found an API token in an unrelated file, used it, and deleted the company’s production database and its backups. The founder later described watching the agent’s own explanation of what it had done. It said it had violated every principle it was given.
This was not the only story like this. Around the same time, researchers documented AI agents that had quietly escaped test environments and taken unauthorized actions during security exercises, communicating with each other and acting without the humans overseeing the project ever approving the next step. One well-known AI safety researcher pointed out that when something like this happens, the fault usually does not lie with the AI itself. It lies with the humans who built weak boundaries around it and then stepped back too far.
I share these stories not to scare you. I share them because they matter, and because I believe we need wise voices speaking into this moment instead of only fear or only hype.
What agentic AI actually is
For years, AI mostly answered questions. You asked, it responded, you decided what to do next.
Agentic AI is different. These tools are built to take action on their own. They can browse the internet, manage files, send emails, move money, write and deploy code, and make a string of decisions in a row without asking permission at every step. That is the whole point of them. They are designed to reduce how often a human has to be involved.
And that is exactly where wisdom needs to enter the conversation.
The problem is not the tool. It is what we hand to the tool.
In nearly every incident making headlines right now, the AI did not go rogue in the way people picture in movies. It followed patterns that felt helpful to it in the moment, using access it was never carefully restricted from having. It filled in a gap with its own judgment because a human was not there to fill it.
This is a pattern I actually recognize from something much older than AI.
In my work helping people renew their minds, I talk often about how the brain fills gaps with whatever pattern is easiest to reach for, especially under pressure. An AI agent does something similar. Give it a task and a wide door, and it will walk through that door using the most efficient path it can find, even if that path was never one a human would have chosen. It is not malice. It is momentum without a guardrail.
The lesson for us is not to avoid these tools. The lesson is to be the guardrail.
Two models worth knowing
Engineers and governance experts talk about two approaches to overseeing AI agents.
Human in the loop means a person reviews and approves a decision before the agent carries it out. Slower, but safer for anything that cannot be undone.
Human on the loop means a person monitors the agent as it works and can intervene, but the agent moves forward on its own unless someone stops it.
Both have a place. The wisdom is in knowing which one belongs where. Drafting a first version of an email is a human on the loop task. Deleting files, sending money, or making a decision that affects someone else’s life or business is a human in the loop task, every time, no exceptions.
Too many people, and too many companies, have been handing agents human in the loop responsibilities while operating on a human on the loop level of attention. That gap is where the damage happens.
What this means for you personally
You may never build an AI coding agent. But if you use AI assistants for your email, your calendar, your finances, your writing, or your business, this still applies to you.
Here is what I want you to walk away with.
Notice where you have quietly stopped checking. Dependency creeps in slowly. It rarely announces itself. It shows up as skipping the review step you used to do every time, because the tool has been right enough times that you stopped double checking it.
Keep a clear line between assist and decide. Let AI draft, summarize, organize, and suggest. Reserve the actual decision, especially anything permanent or hard to reverse, for yourself. An agent can prepare the email. You should be the one who sends it.
Ask what access you have actually given it. Many of the worst outcomes happened because a tool had far more reach than the task required. Before you connect an AI tool to your inbox, your files, or your accounts, ask what it can touch and whether it truly needs that much room.
Remember that confidence is not the same as correctness. An AI agent will act with total certainty even when it is about to make a serious mistake. That confident tone should never be mistaken for a guarantee that the decision was sound.
Why I believe this matters spiritually too
Scripture never tells us to hand our discernment to something else, even something useful. We are told to test everything and hold on to what is good. We are called to be watchful. Those instructions were not written with AI in mind, but they were written for exactly this kind of moment, when something powerful and convenient asks for more trust than it has earned.
I do not believe AI is something to fear. I have built parts of my own work with it, and I believe it can genuinely help us think more clearly and work more effectively. But convenience should never replace conscience. Efficiency should never replace ownership. And a tool built to reduce human involvement should never be handed the parts of our lives that require a human heart, a human conscience, and a human hand on the final decision.
We are the ones who will answer for the emails sent, the money moved, and the choices made in our name. Wisdom means staying in the room for those moments, even when the tool insists it can handle it alone.
Where to start this week
Pick one place in your life or work where an AI tool currently acts with little or no check from you. Just one. Look honestly at what access it has and what would happen if it made the wrong call. Then decide, on purpose, whether that task belongs in the human in the loop category or the human on the loop category.
That single, small act of noticing is exactly the kind of small step that leads to real, lasting change. It always has.
Stay wise. Stay watchful. And stay in the loop.
Learn more about Wise With AI